Meta pushed a new open-weight model into the wild on October 15, 2026, and the headline number isn’t just the parameter count. Muse Spark Open went live simultaneously across seven hosted-inference providers and four separate weight-distribution channels, a rollout pattern that looks less like a research paper drop and more like a coordinated product launch. According to Meta’s own announcement, posted through Meta’s developer blog, the company is calling Muse Spark Open “the most capable US open-weight model to date,” a direct jab at the Chinese labs that have dominated open-weight leaderboards for much of the past year.
The model itself is a 816-billion-parameter mixture-of-experts system with 42 billion parameters active per token, built by Meta Superintelligence Labs. It supports a 1-million-token context window and takes text, images, video, and audio as input, producing text output. Those specs matter, but the bigger story on October 16 is where Muse Spark Open showed up on day one and what that distribution footprint says about how Meta intends to compete for developer mindshare against DeepSeek, Qwen, and GLM.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Meta Announced on October 15
Meta Superintelligence Labs released Muse Spark Open on October 15, 2026, positioning it as a natively multimodal mixture-of-experts model aimed squarely at coding and agentic workloads rather than general chit-chat. The announcement emphasized long-horizon tasks, complex software engineering, and multi-step reasoning as the model’s primary use cases, a framing that puts Muse Spark Open in direct competition with the current generation of agentic coding models rather than consumer chatbots.
Meta’s model card, published through Hugging Face, confirms the architecture detail that makes Muse Spark Open distinctive: 256 experts per mixture-of-experts layer, with 8 selected per token. That routing scheme is what lets a model with 816 billion total parameters run with only 42 billion active per token, keeping inference costs well below what a dense model of equivalent scale would require. It is the same basic design philosophy that has powered the largest open releases out of China over the past 18 months, except this time it is Meta’s name on the model card.
Inside Muse Spark Open’s Architecture
Strip away the marketing language and Muse Spark Open is a straightforward, if enormous, mixture-of-experts transformer. The 816 billion total parameters are split across 256 experts per layer, and the router picks 8 of them for every token that moves through the network. Active parameters per token land at 42 billion, which is the number that actually determines inference speed and GPU memory bandwidth requirements during generation.
The 1-million-token context window is large enough to hold entire codebases, long technical specifications, or multi-document research bundles in a single pass, which lines up with Meta’s stated focus on agentic software engineering. A coding agent that has to re-read a 40-file repository on every turn burns tokens and time; one that can hold the whole thing in context does not. Meta’s framing around “long-horizon tasks” suggests Muse Spark Open is built to stay coherent across many sequential tool calls, not just answer a single prompt well.
Being natively multimodal on the input side, handling text, images, video, and audio, while producing only text output is also a deliberate scoping choice. Meta is not trying to compete with image or video generation models here. It is building a model that can look at a screenshot of a bug, read a video of a reproduction case, or parse a voice note describing a feature request, and then write code or reasoning in response.
Seven Clouds, Four Weight Channels: The Distribution Play
The part of this launch that deserves more attention than it has gotten is the sheer breadth of day-one availability. Muse Spark Open went live through hosted inference on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI simultaneously. Separately, the raw model weights became downloadable through Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models.
Microsoft confirmed its side of that arrangement directly, publishing a post through its own Azure AI Foundry blog stating that Muse Spark Open is available in Microsoft Foundry Models. Intel also moved on day zero, announcing through its own newsroom that it was offering day-0 support for Muse Spark Open, a signal that chipmakers are now racing to validate new frontier-scale open releases the moment they ship rather than weeks later.
That is a meaningfully different launch strategy than how most open-weight releases have historically worked, where a lab drops weights on Hugging Face and waits for the community and infrastructure providers to catch up over following weeks, a pattern that stood in contrast to the pricing-forward launch of Mistral Large 4 earlier this year. Meta appears to have lined up eleven separate distribution partners before the public announcement even went out, which points to a coordinated enterprise sales motion sitting behind what looks, on the surface, like a research release.
| Channel type | Platform | What it offers |
|---|---|---|
| Hosted inference | Baseten | API access to Muse Spark Open |
| Hosted inference | Fireworks AI | API access to Muse Spark Open |
| Hosted inference | Groq | API access to Muse Spark Open |
| Hosted inference | Modal | API access to Muse Spark Open |
| Hosted inference | Nebius | API access to Muse Spark Open |
| Hosted inference | OpenRouter | API access to Muse Spark Open |
| Hosted inference | Together AI | API access to Muse Spark Open |
| Model weights | Amazon SageMaker AI | Downloadable weights for self-hosting |
| Model weights | Google Cloud Model Garden | Downloadable weights for self-hosting |
| Model weights | Hugging Face | Downloadable weights for self-hosting |
| Model weights | Microsoft Foundry Models | Downloadable weights for self-hosting |
Why Running on Two Mac Studios Matters
Meta also demonstrated Muse Spark Open running on a desktop setup built from two Apple M5 Ultra Mac Studios using Apple’s MLX framework, a deployment story covered in depth in our earlier look at the model’s specs and Mac Studio demo. For a model carrying 816 billion total parameters, that is a notable claim. Frontier-scale open models of this size have typically been framed around data-center GPU clusters, not workstation hardware sitting under a desk.
The practical effect, if the claim holds up under independent testing, is that individual developers and small engineering teams could run a model with capabilities Meta is positioning near the frontier without paying for cloud GPU time at all. That changes the calculus for anyone weighing a subscription to a closed, hosted model against the upfront cost of hardware that pays for itself over time. It also reinforces Apple’s pitch that unified memory architecture on Mac Studio hardware is a legitimate platform for serious local AI work, not just a convenience for running smaller, quantized models.
Muse Spark Open vs Muse Wonder: Meta’s Two-Tier AI Strategy
Muse Spark Open is not Meta’s most capable model. That distinction reportedly still belongs to Muse Wonder, the closed, higher-tier system Meta has positioned against OpenAI’s commercial offerings. The split mirrors a strategy Meta’s own leadership has described openly: ship a strong open-weight model to build developer goodwill and ecosystem lock-in, while keeping the most advanced capabilities behind a closed, monetized product.
Alexandr Wang, Meta’s Chief AI Officer, has articulated that logic directly. “Because of both safety and capabilities, we will generally not open-source our most powerful models,” Wang said, according to reporting from TheNextWeb. That statement draws a clean line around Muse Spark Open’s role in Meta’s lineup: it is the company’s strongest card in the open-weight game, but it is deliberately not the best card in Meta’s full hand.
That two-tier approach is becoming the default playbook across the industry. A closed flagship justifies premium pricing and keeps a company’s absolute best capabilities proprietary, while a strong open release builds the kind of developer trust, third-party tooling, and fine-tuning ecosystem that a closed-only strategy cannot. Meta is betting that Muse Spark Open’s reach across seven inference providers and four weight-hosting platforms builds exactly that ecosystem faster than a closed-only approach ever could.
The Open-Weight Race: DeepSeek, Qwen, GLM and Llama
Context matters here. For most of 2025 and into 2026, the open-weight leaderboard conversation has been dominated by Chinese labs. DeepSeek’s mixture-of-experts releases forced a rethink of what “efficient at scale” could mean for open models. Alibaba’s Qwen line built out one of the broadest multilingual and multimodal open catalogs in the industry. Zhipu AI’s GLM models carved out a strong reputation specifically in reasoning and agentic tool use, a rivalry detailed in our recent GLM vs Qwen vs Step benchmark comparison. Meta’s own Llama family, meanwhile, had started to look like the previous generation’s open-weight champion rather than the current one.
Muse Spark Open is Meta’s answer to that shift, and the “most capable US open-weight model to date” framing is a direct claim about reclaiming ground specifically from US-based competitors rather than conceding the open category entirely to labs operating under different regulatory and export conditions. It is worth being precise about what is and isn’t established here: Meta has not published independently reproduced benchmark comparisons against DeepSeek, Qwen, or GLM alongside this release, so any specific performance ranking should be treated as a claim to watch, not a confirmed fact.
| Model family | Developer | Known specs (Muse Spark Open) / general positioning (others) | Context window | Primary focus |
|---|---|---|---|---|
| Muse Spark Open | Meta Superintelligence Labs | 816B total / 42B active parameters, 256 experts per layer, 8 selected per token | 1 million tokens | Coding, agentic reasoning, multimodal input |
| Llama (Meta’s prior open line) | Meta | Specs vary by version; not part of this announcement | Varies by version | General-purpose open ecosystem |
| DeepSeek open models | DeepSeek | Specs vary by version; not part of this announcement | Varies by version | Efficiency-oriented reasoning and coding |
| Qwen open models | Alibaba | Specs vary by version; not part of this announcement | Varies by version | Multilingual, coding, multimodal coverage |
| GLM open models | Zhipu AI | Specs vary by version; not part of this announcement | Varies by version | Reasoning, coding, agent/tool use |
What Industry Data Says About Open-Source AI Adoption
The appetite for open-weight models among enterprise buyers has been building for a while, which helps explain why Meta lined up eleven distribution partners for this launch instead of a quiet Hugging Face drop. Research from McKinsey & Company found that “more than 50% of respondents reported leveraging open source in each of the data, models, and tools areas of the technology stack,” underscoring that open-weight adoption is no longer a fringe engineering preference but a mainstream procurement pattern.
The same research found that “nearly three-quarters of respondents, or more than 70%, said they were open to either open-source or proprietary technologies across areas of the technology stack,” which suggests most enterprise technology buyers are pragmatic rather than ideological about open versus closed models, a dynamic that favors whichever lab ships the most accessible, best-distributed option at a given moment. Notably, McKinsey also found that among respondents using open-source models, “61% had used Meta’s Llama, compared with 40% who had used Google’s Gemma,” which gives Meta a real incumbency advantage heading into the Muse Spark Open launch: a large share of the enterprise open-model user base already has existing Llama tooling and pipelines that could plausibly carry over.
Market Impact and Enterprise Cloud Adoption Signals
The clearest market signal from this launch isn’t a stock price move, it is the infrastructure commitments made before the announcement even landed. When Amazon, Google, Microsoft, and Hugging Face all agree to host the same model’s weights on day one, and Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI all agree to serve it through hosted APIs simultaneously, that represents a meaningful amount of engineering and business-development coordination across companies that compete with each other in other contexts.
Intel’s day-zero support announcement adds another layer. Chipmakers validating a new open-weight release immediately, rather than weeks after community benchmarking begins, signals that hardware vendors see enough demand likely to justify the engineering cost of early optimization work. South Korea’s Chosun Ilbo covered the launch as a joint Meta-Apple moment, framing the unveiling around both companies’ involvement, which points to how the local-hardware angle is resonating outside the US tech press as well.
Historical Context: From Llama to Muse Spark Open
Meta’s open-weight strategy did not start with Muse Spark Open. The company built its open-model reputation on the Llama family, which became the default starting point for countless fine-tuned derivatives, enterprise deployments, and academic research projects over the past several years. That head start is reflected in the McKinsey adoption numbers cited above, where Llama’s usage rate among open-source adopters dwarfed the next closest competitor tracked in that research.
What changed over the past 18 months is that Chinese labs caught up and, by many community accounts, pulled ahead on headline open-weight benchmarks, particularly in coding and reasoning. Muse Spark Open represents Meta’s attempt to reset that narrative, pairing a genuinely large mixture-of-experts architecture with the kind of day-one, multi-platform distribution that Llama releases didn’t always get. It is a bet that distribution breadth, not just raw capability, is what actually drives adoption in a market now crowded with credible open alternatives, including efforts like Switzerland’s fully open Apertus 1.5 and Reflection AI’s open-weight bet against China’s open models.
Licensing: Open-Weight Does Not Mean Unrestricted
One distinction worth flagging for developers evaluating Muse Spark Open: open-weight and fully permissive open-source are not the same thing. Coverage of the release, including reporting from OkayNews on the model’s commercial licence conditions, indicates that Meta’s distribution terms attach specific usage and commercial conditions to the downloadable weights. That is consistent with how Meta has historically licensed its open-weight model families, and it means engineering and legal teams evaluating Muse Spark Open for commercial deployment should read the license terms directly rather than assume unrestricted use, regardless of how the model is marketed.
Predictions: What Comes Next for the Open-Weight Race
A few trends look likely to play out over the coming months as Muse Spark Open settles into the market. First, expect independent benchmark comparisons against DeepSeek, Qwen, and GLM to surface within weeks, since the developer community that uses OpenRouter and Together AI tends to run its own head-to-head evaluations quickly once a model is broadly accessible through hosted APIs.
Second, the two-Mac-Studio demonstration will likely push more labs to explicitly test and market local-hardware deployment paths for their largest open releases, especially as Apple’s MLX framework matures and unified-memory configurations get larger. Third, expect other major labs to respond to Meta’s multi-platform, multi-cloud distribution strategy by lining up similarly broad day-one partner lists for their own next releases, since a narrow, single-platform open-weight drop increasingly looks like a competitive disadvantage.
Fourth, enterprise buyers already running Llama-based pipelines are likely to be the fastest adopters of Muse Spark Open, given the incumbency pattern McKinsey’s research highlights. Fifth, the gap between Muse Spark Open and Muse Wonder will likely remain a talking point, with developers and analysts continuing to probe exactly how much capability Meta is withholding from its open tier versus its closed, monetized flagship.
Risks and Unanswered Questions
Several questions remain genuinely open. Meta has not published independently verifiable benchmark scores alongside this release, so claims about where Muse Spark Open ranks against DeepSeek, Qwen, or GLM remain unconfirmed pending third-party testing. The exact commercial license terms, beyond the general existence of commercial conditions reported by outlets covering the release, have not been fully detailed in Meta’s own announcement materials reviewed for this article. And while the two-Mac-Studio demonstration is a notable claim, real-world throughput and usability for that configuration under sustained workloads has not yet been independently measured outside of Meta’s own demonstration.
There is also the open question of how long Meta sustains this level of multi-platform coordination for future releases. Lining up eleven distribution partners for a single launch is a heavy lift, and whether that becomes Meta’s standing operating model for every future open-weight drop, or a one-time push specifically designed to win back developer attention from Chinese labs, will become clearer with the next release in this line.
Frequently Asked Questions
What is Muse Spark Open?
Muse Spark Open is an open-weight, natively multimodal mixture-of-experts AI model released by Meta Superintelligence Labs on October 15, 2026. It carries 816 billion total parameters with 42 billion active per token and supports a 1-million-token context window.
Where can I access Muse Spark Open?
Hosted inference is available through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Downloadable model weights are available through Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models.
Is Muse Spark Open the same as Muse Wonder?
No. Muse Wonder is reported to be Meta’s more advanced, closed model positioned against commercial offerings from other AI labs. Muse Spark Open is Meta’s open-weight release and is not described as Meta’s most capable system overall.
Can Muse Spark Open really run on a Mac Studio?
Meta demonstrated the model running across two Apple M5 Ultra Mac Studios using Apple’s MLX framework. Independent, third-party verification of sustained real-world performance on that setup has not yet been widely published.
Is Muse Spark Open fully open source?
It is described as open-weight, meaning the model weights are downloadable, but reporting on the release indicates the weights carry specific commercial licensing conditions. That distinguishes it from a fully permissive open-source software license.
How does Muse Spark Open compare to DeepSeek, Qwen, and GLM?
Meta has not published independently reproduced benchmark comparisons against those model families alongside this release. Muse Spark Open’s architecture, 256 experts per layer with 8 selected per token, is broadly similar in design philosophy to recent large open releases from those labs, but specific performance rankings remain unconfirmed.
What is Muse Spark Open primarily designed for?
Meta’s announcement frames the model around coding and agentic work, including long-horizon tasks, complex software engineering, and multi-step reasoning, rather than general-purpose consumer chat use cases.
Why did so many cloud and inference providers support Muse Spark Open on day one?
Eleven separate platforms, spanning hosted inference and weight distribution, went live simultaneously with the announcement, suggesting Meta coordinated a broad distribution rollout ahead of the public release rather than relying on gradual community adoption.
